A unified multi-level spectral–temporal feature learning framework for patient-specific seizure onset detection in EEG signals

Epileptic seizure onset detection in electroencephalography (EEG) signals is a challenging task due to the severe variation of seizures. Recently, automatic seizure onset detection frameworks fail to fully consider both nonstationary and stochastic characteristics of EEGs in nature, which may lead to information default and further produce suboptimal recognition performance consequently. In this work, we propose a patient-specific seizure onset detection method based on fully exploration of auxiliary supplementary spectral–temporal information in EEG signals. Specifically, prior to feature extraction procedure, EEG signals are firstly decomposed into 5 groups of coefficients at different levels based on the clinical interest. Representative feature in temporal-domain, which is a translation of the nonlinear property of EEG signals, is then extracted by a combination of principal component analysis and common spatial pattern (PCA-CSP) and multivariate multiscale sample entropy (MMSE) in parallel and dimensionally reduced by a tree-based feature selection algorithm. Supplementary information in spectral-domain is further explored by the proposed unified maximum mean discrepancy autoencoder (uMMD-AE). Finally, an optimal combination of features above is identified and fed into a series of support vector machine classifiers with a decision fusion module for the intelligent recognition of epileptic EEGs. The proposed method achieves an average sensitivity, latency and false detection rate of 97.2%, 1.10s and 0.64/h respectively on Children Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) Scalp EEG Database. Competitive experimental results demonstrate the efficacy of the proposed unified multi-level spectral–temporal feature learning framework in epileptic EEG recognition, validating its effectiveness in the automatic patient-specific seizure onset detection.